Survival Analysis with Accelerated Failure Time
What is survival analysis?
Survival analysis (regression) models time to an event of interest. Survival analysis is a special kind of regression and differs from the conventional regression task as follows:
The label is always positive, since you cannot wait a negative amount of time until the event occurs.
The label may not be fully known, or censored, because “it takes time to measure time.”
The second bullet point is crucial and we should dwell on it more. As you may have guessed from the name, one of the earliest applications of survival analysis is to model mortality of a given population. Let’s take NCCTG Lung Cancer Dataset as an example. The first 8 columns represent features and the last column, Time to death, represents the label.
Inst 
Age 
Sex 
ph.ecog 
ph.karno 
pat.karno 
meal.cal 
wt.loss 
Time to death (days) 

3 
74 
1 
1 
90 
100 
1175 
N/A 
306 
3 
68 
1 
0 
90 
90 
1225 
15 
455 
3 
56 
1 
0 
90 
90 
N/A 
15 
\([1010, +\infty)\) 
5 
57 
1 
1 
90 
60 
1150 
11 
210 
1 
60 
1 
0 
100 
90 
N/A 
0 
883 
12 
74 
1 
1 
50 
80 
513 
0 
\([1022, +\infty)\) 
7 
68 
2 
2 
70 
60 
384 
10 
310 
Take a close look at the label for the third patient. His label is a range, not a single number. The third patient’s label is said to be censored, because for some reason the experimenters could not get a complete measurement for that label. One possible scenario: the patient survived the first 1010 days and walked out of the clinic on the 1011th day, so his death was not directly observed. Another possibility: The experiment was cut short (since you cannot run it forever) before his death could be observed. In any case, his label is \([1010, +\infty)\), meaning his time to death can be any number that’s higher than 1010, e.g. 2000, 3000, or 10000.
There are four kinds of censoring:
Uncensored: the label is not censored and given as a single number.
Rightcensored: the label is of form \([a, +\infty)\), where \(a\) is the lower bound.
Leftcensored: the label is of form \([0, b]\), where \(b\) is the upper bound.
Intervalcensored: the label is of form \([a, b]\), where \(a\) and \(b\) are the lower and upper bounds, respectively.
Rightcensoring is the most commonly used.
Accelerated Failure Time model
Accelerated Failure Time (AFT) model is one of the most commonly used models in survival analysis. The model is of the following form:
where
\(\mathbf{x}\) is a vector in \(\mathbb{R}^d\) representing the features.
\(\mathbf{w}\) is a vector consisting of \(d\) coefficients, each corresponding to a feature.
\(\langle \cdot, \cdot \rangle\) is the usual dot product in \(\mathbb{R}^d\).
\(\ln{(\cdot)}\) is the natural logarithm.
\(Y\) and \(Z\) are random variables.
\(Y\) is the output label.
\(Z\) is a random variable of a known probability distribution. Common choices are the normal distribution, the logistic distribution, and the extreme distribution. Intuitively, \(Z\) represents the “noise” that pulls the prediction \(\langle \mathbf{w}, \mathbf{x} \rangle\) away from the true log label \(\ln{Y}\).
\(\sigma\) is a parameter that scales the size of \(Z\).
Note that this model is a generalized form of a linear regression model \(Y = \langle \mathbf{w}, \mathbf{x} \rangle\). In order to make AFT work with gradient boosting, we revise the model as follows:
where \(\mathcal{T}(\mathbf{x})\) represents the output from a decision tree ensemble, given input \(\mathbf{x}\). Since \(Z\) is a random variable, we have a likelihood defined for the expression \(\ln{Y} = \mathcal{T}(\mathbf{x}) + \sigma Z\). So the goal for XGBoost is to maximize the (log) likelihood by fitting a good tree ensemble \(\mathcal{T}(\mathbf{x})\).
How to use
The first step is to express the labels in the form of a range, so that every data point has two numbers associated with it, namely the lower and upper bounds for the label. For uncensored labels, use a degenerate interval of form \([a, a]\).
Censoring type 
Interval form 
Lower bound finite? 
Upper bound finite? 

Uncensored 
\([a, a]\) 
✔ 
✔ 
Rightcensored 
\([a, +\infty)\) 
✔ 
✘ 
Leftcensored 
\([0, b]\) 
✔ 
✔ 
Intervalcensored 
\([a, b]\) 
✔ 
✔ 
Collect the lower bound numbers in one array (let’s call it y_lower_bound
) and the upper bound number in another array (call it y_upper_bound
). The ranged labels are associated with a data matrix object via calls to xgboost.DMatrix.set_float_info()
:
import numpy as np
import xgboost as xgb
# 4by2 Data matrix
X = np.array([[1, 1], [1, 1], [0, 1], [1, 0]])
dtrain = xgb.DMatrix(X)
# Associate ranged labels with the data matrix.
# This example shows each kind of censored labels.
# uncensored right left interval
y_lower_bound = np.array([ 2.0, 3.0, 0.0, 4.0])
y_upper_bound = np.array([ 2.0, +np.inf, 4.0, 5.0])
dtrain.set_float_info('label_lower_bound', y_lower_bound)
dtrain.set_float_info('label_upper_bound', y_upper_bound)
library(xgboost)
# 4by2 Data matrix
X < matrix(c(1., 1., 1., 1., 0., 1., 1., 0.),
nrow=4, ncol=2, byrow=TRUE)
dtrain < xgb.DMatrix(X)
# Associate ranged labels with the data matrix.
# This example shows each kind of censored labels.
# uncensored right left interval
y_lower_bound < c( 2., 3., 0., 4.)
y_upper_bound < c( 2., +Inf, 4., 5.)
setinfo(dtrain, 'label_lower_bound', y_lower_bound)
setinfo(dtrain, 'label_upper_bound', y_upper_bound)
Now we are ready to invoke the training API:
params = {'objective': 'survival:aft',
'eval_metric': 'aftnloglik',
'aft_loss_distribution': 'normal',
'aft_loss_distribution_scale': 1.20,
'tree_method': 'hist', 'learning_rate': 0.05, 'max_depth': 2}
bst = xgb.train(params, dtrain, num_boost_round=5,
evals=[(dtrain, 'train')])
params < list(objective='survival:aft',
eval_metric='aftnloglik',
aft_loss_distribution='normal',
aft_loss_distribution_scale=1.20,
tree_method='hist',
learning_rate=0.05,
max_depth=2)
watchlist < list(train = dtrain)
bst < xgb.train(params, dtrain, nrounds=5, watchlist)
We set objective
parameter to survival:aft
and eval_metric
to aftnloglik
, so that the log likelihood for the AFT model would be maximized. (XGBoost will actually minimize the negative log likelihood, hence the name aftnloglik
.)
The parameter aft_loss_distribution
corresponds to the distribution of the \(Z\) term in the AFT model, and aft_loss_distribution_scale
corresponds to the scaling factor \(\sigma\).
Currently, you can choose from three probability distributions for aft_loss_distribution
:

Probability Density Function (PDF) 


\(\dfrac{\exp{(z^2/2)}}{\sqrt{2\pi}}\) 

\(\dfrac{e^z}{(1+e^z)^2}\) 

\(e^z e^{\exp{z}}\) 
Note that it is not yet possible to set the ranged label using the scikitlearn interface (e.g. xgboost.XGBRegressor
). For now, you should use xgboost.train
with xgboost.DMatrix
.